July 21, 2017 at 10:12 am #2159
SWICCA recently started offering seasonal forecasts of climate and hydrological indicators using a similar user interface as for the more long term climate change related indicators. Once a month (around the 10th), probabilistic forecasts are presented for riverflow, precipitation and temperature for 7 months into the future. Maps show information on the likelyhood of anomalous values of the indicators, i.e. indicating whether they will be higher than, lower than, or close to the long term average (AN, BN and NN, resp.). This information can (and should) be combined with information on the forecast skill of each of these indicators. For particular basins, the numbers ‘behind the maps’ can be downloaded in the form of monthly values of that indicator for each of the 51 ensemble members contributing to the forecast.
Yet, since this is a very new data product we realize that our purveyors may have questions as to how to best use this information. Therefore, we set up this topic. We challenge the users to first briefly describe how they intend to use these seasonal forecasts with their clients. Next, you can post ANY questions you may have encountered when using seasonal forecast and we will try to answer or at least reflect on them. At a later stage we can use these discussions to develop a more tailor made tutorial.
By sharing our questions and answers, concerns and enthusiasm we hope you come to appreciate the value this type of information may potentially have for you and your clients.July 28, 2017 at 2:10 pm #2193
The intention of UPV is to use seasonal forecasts though a modelling chain to advise in advance our client and the interested parties about a better way to manage water resources and its allocation in different uses for the coming months, taking into account the probabilities of having more or less water resources in the basin. For example, in the case of agriculture these predictions could be valuable for decisions about the type of crops to grow in the season and the need of insurances, if is expected a period of water stress for the crops. Thus, the aim of this analysis is to provide reliable estimations to our client in order to advice all users about the procedure in a water stress or drought scenario.
However, in the exploration of the Maps interface of Seasonal Forecasts, when we remove results without skill in precipitation, temperature and river flows, the area not masked is very small in the Jucar River Basin, this means that the area without skill in this basin is very big. This leads us to think about the reliability of the seasonal forecast in this area. If we are right, the lack of skill could lead to not trustable results. So, we would like to know the opinion of data providers and the other project members about this issue. We will appreciate any suggestion and feedback you may have for improving our plan.August 3, 2017 at 9:52 am #2213
This is an excellent point. The aim of the masks is for users to be aware of the past performance of the system in each catchment, with reference to climatology. By climatology, we mean a forecast ensemble based on past model simulations. Here, the masks are produced based on an evaluation criterion which assesses simultaneously the reliability of the forecasts, their confidence, as well as their uncertainty (CRPS).
Based on this, if, for a given month of the year and lead time, parts of the Jucar River Basin are masked, we can indeed not guarantee that the produced forecasts are reliable, because historically, climatology has performed better.
Please note that masks differ for each month of the year and each lead time. Therefore, even if parts of the Jucar may be masked in July, they may not be masked in August.
Note also that streamflow forecasts can perform well even when meteorological variables have no skill, due to the slow catchment processes.August 7, 2017 at 8:09 am #2217
Dear Sara, Louise,
indeed skill in forecasting streamflow is varies (a lot) with lead time, time of the year and river basin. And the forecast quality for discharge can be better than that for precipitation. For Spain we know from previous work (NOT this EHYPE model, but statistical models) that streamflow forecast quality is quite good for at least Central, Western Spain (Douro, Tejo, Guadiana) for winter and early spring (DJFM). Also our own model shows potential skill in this period and basins, but also in july/august more towards your area (Mediterranean basins).
Another issue is the used metric for the skill. The maps presented are masked out based on CPRS, this is a measure across all flow percentiles and therefore perhaps a bit strict. Other skill scores that measure the performance of forecasting only for high flow or low flows, such as ROCS may give less conservative indications of forecast quality. Refering the previous examples: in winter high flows are better predicted than low flows, in summer the other way around: low flows are better predicted than high flows. Moreover, more extreme anomalies are often better forecasted than less extreme ones.
May be Louise can comment a bit more already on winter skill and skill for high or low flows for the EHYPE model.August 17, 2017 at 11:48 pm #2242
Maria J. PoloParticipant
The UCO team develops the case study “Snow effects for water availability” in the Guadalfeo River Basin, and these seasonal forecasts can be really valuable for our clients by using the river flow forecast as a basis for water resource management; additionally, the combination of precipitation and temperatura forecasts can assess the probability of snow occurrence and, thus, the likely seasonal pattern of the streamflow regime for the next months. The Rules Reservoir operational system particularly will be the primary user for this new service, followed by the small hydropower facilities (exploiting medium and low flows associated to both the wet and the snowmelt seasons) in the northern area of the basin. We, as knowledge purveyors, are first testing the seasonal forecasts provided so far (from May on) against the later observations, and checking their local skill. So far, river flow forecasts have provided very good performance in our basin, which is really promising; weather variables are less skilled, especially temperature has been overestimated, which may be a constraint since snow occurrence cannot be then accurately forecast from these data.
We need to wait for wet season to check out these preliminary results also during the medium and high flow periods. The reliability of the forecasts is key for getting the users involved in using them and getting advantage from them. Following this, we find really interesting the possibility of including different metrics to assess the skill of the forecasts, since we fully agree with the last comment posted.August 23, 2017 at 8:30 am #2253
There is a lot of interest in seasonal forecasts among hydropower producers. Companies in Sweden and Norway are mainly interested in seasonal forecasts for the snow melt period in spring/summer. However there are still many questions about skill and many hydropower companies still rely on historical records (climatology) to generate long-time scenarios for production planning. One question that has been discussed is if the skill of seasonal forecasts in some way could be linked to present weather situation on a larger scale? For example, if seasonal forecasts (for a given region) can be found to produce better during wet/cold/dry years? That sort of information might be useful for hydropower producers as a first step in starting to use seasonal forecasts.August 31, 2017 at 12:17 pm #2365
I am happy to hear the river flow forecasts performed well so far. You have to realize that the influence of snow accumulation in winter, and snow melting in spring/summer, is explicitly accounted for in our forecast. In fact it is a very important source of the skill we see. Based on precipitation and temperature (corrected for altitude, see below) the model continuously computes the thickness of the snowpack (in so-called snow-water equivalents). Based on that we have a fairly good idea how much snow there is in any basin at the start of a forecast. This so-called inital condition provides so much ‘memory’ to the system that it allows the model to predict snow melting and therefore river flow in spring/early summer fairly well. In fact the (poor) quality of the precipitation forecast becomes less important, especially when -like in your basin- in the melting season the rainfall is relatively low.
Further you mention that temperatures are overestimated. I dont know compared to what reference, but of course temperature is very sensitive to altitude. So if you compare predicted temperatures to some mountain weather station data, there will generally be a systematic bias. A bias that is fairly constant and that can easily be corrected, if we know the difference between station altitude and model terrain elevation. Having said that temperature anomaly forecasts are generally much better than precipitation anomaly forecasts, I would expect so too in you region. Note that I say ‘anomalies’: they are by definition not senstitive to systematic errors!
Finally, to connect what I said in the previous two paragraphs: for snow accumulation and melt our models work with altitude bands. So even if your basin is represented by only a few model units (sub basins in the EHYPE, grid cells in may other models), within each unit the models keep track of different altitudes with each a different temperature to account for the unresolved topography of the mountains.August 31, 2017 at 12:53 pm #2366
I’d like to respond to two issues you raise:
– the skill of predicting snow melt in spring summer. As I also explained in previous post (just above this one), the models do this fairly well primarily because we know the amount of snow present in a basin at the start of a forecast also fairly well. So if you mean by ‘linking forecast’s to present large scale weather’ linking it to present large scale snow distribution, then that is exactly what the forecast models do!
– however if you mean linking it to the present atmospheric state, that is not something we do (yet), but that might be a good idea to explore and do in the future. Then we would be issueing so-called conditional forecasts. For example, if in December we are in a specific strong positive or negative state of the North Atlantic Oscillation (NAO, an important predictor of European weather), then we might have more confidence in our forecasts than if the NAO is weakly developed.
To some extend this is done implicitly in our model-chain as our maps show high, medium, more low probabilities of e.g. Above Normal rain or river flow. A high probability means we are more confident about the forecast. However, at this stage we cannot tell you why (in causal terms) it is high, we only know then that the majority of the model runs predict this. More research (and operational development) is needed before we can tell you for each forecast why high or low river flows are predicted.August 31, 2017 at 4:11 pm #2367
Dear Linnéa, Maria and Ronald,
I will quickly add a few comments to what Ronald already wrote.
Maria, thank you very much for this feedback. It is extremely valuable to hear about local model performance. I am looking forward to hearing your feedback for wetter months, once the time comes. As for temperature and precipitation, note that a bias correction based on the GFD dataset has now been implemented (as of August 25th). Hopefully bias corrected forecasts will better fit local observations. I really like the suggestion of adding other evaluation metrics, and it could be explored in the future. It does raise the challenge of graphically communicating large amounts of data while keeping clear the purpose of each metric, and to overcome this, user feedbacks would be key.
Linnéa, I also really like the idea to link seasonal forecast skill with wetter/colder/drier conditions. For now, the skill graphs in the interface provide this information for each month of the year. The month of the year can be seen as a proxy to these changing conditions, if you consider the entire year. For a more detailed evaluation, of, for example, dry months of January, longer datasets would be necessary to obtain statistically robust evaluations.
The forum ‘SWICCA Forum’ is closed to new topics and replies.